HYZ-01-0.6B-Base/README.md

---
language:
- tr
- en
license: apache-2.0
tags:
- turkish
- turkish-llm
- turkish-nlp
- base-model
- neuroturk
- hyz01
- text-generation
- cpt
- qwen3
- 4-bit-precision
- gguf
- quantized
- pytorch
library_name: transformers
pipeline_tag: text-generation
---

<div align="center">


**Open-source base language model pre-trained for Turkish by NeuroTürk**

![Ekran_görüntüsü_2026-05-17_141827-removebg-preview](https://cdn-uploads.huggingface.co/production/uploads/69ee4bc4e6c1cd02a169fcd2/8NC0K713fHf9g8zQqZHbs.png)

[![License](https://img.shields.io/badge/License-Apache%202.0-blue)](LICENSE)
[![Language](https://img.shields.io/badge/Primary%20Language-Turkish-red)]()
[![HuggingFace](https://img.shields.io/badge/🤗-NeuroTurk-yellow)](https://huggingface.co/NeuroTurk)
[![GitHub](https://img.shields.io/badge/-NeuroTürk-181717?logo=github&logoColor=white)](https://github.com/neuroturk)
[![X](https://img.shields.io/badge/-NeuroTürk-000000?logo=x&logoColor=white)](https://x.com/neuroturk_ai)


</div>

---

## 1. Introduction

HYZ-01-0.6B-Base is the **base (pre-trained only) version** of the HYZ-01 series developed by **NeuroTürk**. It is a raw language model that has undergone multi-stage Turkish continual pre-training (CPT) on top of a multilingual foundation, without any instruction tuning or alignment. It is intended for researchers and developers who wish to fine-tune the model for their own tasks.

The model is built on a multilingual foundation covering 119 languages and has been continuously pre-trained with a focus on Turkish. The tokenizer has been extended specifically for Turkish morphological structure and advanced use cases. **HYZ-01-0.6B-Base is the lightweight, open-source base version of HYZ-01, developed by NeuroTürk for Turkish.**

> Note: This is the base pre-trained version. For the instruction-tuned version, see: [HYZ-01-0.6B](https://huggingface.co/neuroturk/HYZ-01-0.6B)

---

## 2. Model Summary

### Continual Pre-Training

- **Base model:** 4-stage Turkish continual pre-training (CPT) applied on top of a multilingual foundation.
- Training stages include general Turkish web corpus, curated domain data, Wikipedia, and high-quality filtered text.
- Optimization: bfloat16, flash-attention-2, AdamW.

### Tokenizer Extension

New special tokens were added to the tokenizer for two purposes:

- **Language-structure tokens:** To represent Turkish morphological features more efficiently.
- **Task and structure tokens:** To support structural use cases such as chain-of-thought, code blocks, section markers, and language labels.

The following 20 tokens have been added to the vocabulary and are reserved as infrastructure for future advanced capabilities:

| Group | Tokens | Future Use |
|:---|:---|:---|
| Brand | `<\|neuroturk\|>` `<\|hyz01\|>` `<\|tr\|>` `<\|en\|>` | Model identity and multilingual control |
| Chain-of-Thought | `<\|think\|>` `<\|/think\|>` `<\|step\|>` `<\|answer\|>` | Step-by-step reasoning (CoT) |
| Dialogue | `<\|system\|>` `<\|user\|>` `<\|assistant\|>` `<\|end\|>` | Multi-turn dialogue and role management |
| Code | `<\|code\|>` `<\|/code\|>` `<\|output\|>` `<\|error\|>` | Structured code generation and debugging |
| Structure | `<\|title\|>` `<\|section\|>` `<\|list\|>` `<\|note\|>` | Long-form and structured text generation (reports, articles, etc.) |

---

## 3. Model Details

| Feature | Value |
|:---|:---|
| Total parameters | 595,798,016 (~0.6B) |
| Non-embedding parameters | 440,467,456 (~0.44B) |
| Hidden dimension | 1,024 |
| Number of layers | 28 |
| Attention heads (Q) | 16 |
| Attention heads (KV) | 8 (GQA) |
| Head dimension | 128 |
| Activation | SiLU |
| Normalization | RMSNorm (ε = 1 × 10⁻⁶) |
| Positional encoding | RoPE (θ = 1,000,000) |
| Vocabulary size | 151,690 |
| Training context length | 4,096 tokens |
| Theoretical max context | 32,768 tokens |
| Precision | BFloat16 |
| VRAM usage (fp16) | ~1.11 GB |
| Disk size | ~1.11 GB |

---

## 4. Training Details

| Setting | Value |
|---|---|
| Training type | Continual Pre-Training (CPT) |
| Number of stages | 4 |
| Optimization | AdamW |
| Precision | BFloat16 |
| LR schedule | Cosine with warmup |
| Context length | 4,096 tokens |

---

## 5. Usage

> **Warning:** This is a base model. It is not instruction-tuned and will not follow instructions reliably. For conversational or task-oriented use, use the instruction-tuned version: [HYZ-01-0.6B](https://huggingface.co/neuroturk/HYZ-01-0.6B)

### Installation

```bash
pip install transformers torch accelerate
```

### Text Generation (Completion)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "neuroturk/HYZ-01-0.6B-Base"

tokenizer = AutoTokenizer.from_pretrained(
    model_name,
    trust_remote_code=True,
    fix_mistral_regex=True 
)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = "Yapay zeka, bilgisayar sistemlerinin"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=200,
    temperature=0.8,
    top_p=0.95,
    do_sample=True,
    repetition_penalty=1.1,
)


new_tokens = outputs[0][inputs['input_ids'].shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))
```

### Low VRAM (4-bit Quantization)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
)

tokenizer = AutoTokenizer.from_pretrained(
    "neuroturk/HYZ-01-0.6B-Base",
    trust_remote_code=True,
)
model = AutoModelForCausalLM.from_pretrained(
    "neuroturk/HYZ-01-0.6B-Base",
    quantization_config=bnb_config,
    device_map="auto",
)
```

### Fine-Tuning with Unsloth

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="neuroturk/HYZ-01-0.6B-Base",
    max_seq_length=4096,
    load_in_4bit=True,
)

model = FastLanguageModel.get_peft_model(
    model,
    r=32,
    lora_alpha=64,
    lora_dropout=0.0,
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ],
    use_gradient_checkpointing="unsloth",
)
```
---

### GGUF Quantizations

For faster inference and lower resource usage, GGUF quantized versions of HYZ-01-0.6B-Base are available. These were kindly provided by [mradermacher](https://huggingface.co/mradermacher).

You can find them here: [__HYZ-01-0.6B-Base-GGUF__](https://huggingface.co/mradermacher/HYZ-01-0.6B-Base-GGUF)

**Using with llama.cpp**

1. Download the GGUF file (e.g., `hyz-01-0.6b-base-q4_k_m.gguf`) from the repository above.
2. Run with `llama.cpp`:
   ```bash
   ./main -m hyz-01-0.6b-base-q4_k_m.gguf -p "Your prompt here" -n 512
For a detailed explanation of quantization types (e.g., Q4_K_M, Q5_K_M), see the llama.cpp documentation.

> Note: These GGUF files are not officially maintained by NeuroTürk, but they are community-tested and widely used. Thanks again to mradermacher for the contribution.

---

## 6. Limitations

- This is a base model without instruction tuning — it will not follow instructions reliably.
- Complex multi-step reasoning may be limited with 0.6B parameters.
- Biases present in the training data may be reflected in outputs.
- Performance drops significantly in languages other than Turkish.
- Human verification of outputs is recommended for critical applications.

---

## 7. Citation

```bibtex
@misc{neuroturk2026hyz01,
  author       = {NeuroTürk},
  title        = {HYZ-01-0.6B: A Lightweight Turkish Base Model},
  year         = 2026,

}
```

---

<div align="center">
<sub>NeuroTürk · HYZ01 · 2026</sub>
</div>